The alignment problem is an extremely difficult one, and one that we've basically punted on for all AI research. Making an algorithm is easy, scoring content by

The alignment problem is an extremely difficult one, and one that we've basically punted on for all AI research. Making an algorithm is easy, scoring content by external factors like engagement or time on screen are also easy. Designing a scoring algorithm that is aligned to a person's preferences, desires, opinions, etc are extremely hard. At best we attempt to find proxies and correlations, the algorithm tracks engagement and we assume that means they want to see more similar content. IMO this is the fundamental reason for chronological feeds. Aligning the algorithm is an unsolved problem, I'd rather just not use it at all.