“The AI safety field should focus on solving the alignment problem rather than just trying to control misaligned AI.”