Introduction to the Alignment Problem in ML and AI

About the talk

When you train an RL model, you have to specify an objective. But can gradient descent find optimizers for something different from the objective you gave? What happens if this objective doesn’t perfectly represent what you want?

This talk will cover

• Mesa-optimizers in RL and what can be done about them
• Why it’s hard to align the objectives of advanced ML systems pursue with what we want
• Approaches used for current SOTA ML models
• The alignment problem in advanced AI systems