Gradient
Lesson 10.10ConceptCode~50 min

Safety & alignment

Getting models to do what we actually mean: reward hacking, calibration, looking inside the network, and fairness.

Module 10 · 0/11 done

Doing what we actually mean

Every model in this course optimised a number you chose: a loss, a reward, an accuracy. Optimisers are relentless and literal. They find whatever raises the number, including ways you didn’t intend. Alignment is the work of making sure a capable system pursues what we actually want. Safety is the broader work of making sure it doesn’t cause harm along the way.

This isn’t a far-future worry. It shows up in today’s systems, in small, measurable ways, and you have the tools to measure them.