AI alignment problem: Mesa-optimizers and inner alignment

Copying my comment from the last time I posted a Robert Miles video:

This video works best when viewed as part of his channel as a whole, which is dedicated to explaining AI safety, starting with the assumption of little to no technical knowledge. These are a bit of a continuation of his videos for Computerphile which lay some groundwork and address common misconceptions or pop culture ideas for solving AI safety problems (with some more clickbaity titles tacked on). His most worthwhile work is probably the multi-parter explaining the Concrete Problems in AI Safety paper for a lay audience.