AI Safety and Alignment
The big idea: a modern AI system does exactly what its objective tells it to do — not what you were secretly hoping it would do. AI safety is the whole engineering effort of making systems behave the way we intend, without harmful surprises. Alignment is the harder inner problem of getting the goal itself right, so that "do what the objective says" and "do what we actually wanted" become the same thing.
This is not science fiction and it is not doom. It is a practical engineering discipline, already visible in the AI products you use every day. By the end of this lesson you will understand why smart people worry about goals, what real safety work looks like inside a chatbot, and how to tell a genuine near-term risk from a speculative long-term one — calmly, with evidence, and without hype.
Learning goals
- Define safety and alignment, and explain how they differ.
- Explain specification gaming: why an optimizer does what you said, not what you meant — using real logged examples.
- Show why "just write better rules" fails, by trying to fully specify "clean the room safely."
- State Goodhart's law and give both an everyday and an AI example.
- Describe instrumental goals (resources, self-preservation) and why they worry researchers — without any "evil AI" story.
- Name the present-day safety tools inside real products (RLHF, red teaming, refusal training, content filters, interpretability, evals) and each one's limits.
- Separate near-term concrete risks from speculative long-term concerns, and describe the honest state of the debate.
- Explain why safety is hard engineering, and what responsible practice (staged releases, audits, incident reporting) looks like.