I've been building a search engine for lectures (https://www.findlectures.com https://www.findlectures.com), so I've spent some time comparing search results vs

I've been building a search engine for lectures (https://www.findlectures.com https://www.findlectures.com), so I've spent some time comparing search results vs. youtube. They seem to weight popularity heavily, which makes sense if you're trying to sell ads, but it means a lot of toxic content rises to the top (e.g. conspiracy theories, or tons of short videos denouncing the video of a well known figure, which becomes un-findable). I'm banking on being able to improve this for my niche use case with more quality measures, although time will tell if this pans out. E.g. if a speaker has spoken at prestigious conferences, written a book, what Wikipedia says about that, or that they've invested in better production quality, including closed captions, better audio/video. For the filter bubble problem, I think that being able to see lists of talks without the machine learning what you "want" is potentially a superior experience. I suspect that if you start by filtering to just the higher quality material, you can potentially get a good cross-section of different viewpoints, just by using text text analysis.