AI researcher Delip Rao notes embeddings-based similarity as a core mechanism in recommendation systems for measuring semantic closeness between items or users
Leshem Choshen adds that metrics duplicate audience exposure for researchers and influencers.
yes, the problem is AI researchers are coveting a large overlapping subset of the audience AI influencers are. In the past, that was not a thing. AI researchers cared about other AI researcher and traditional media attention. With death of trad media, incentives beyond traditional academic incentives, current AI researchers are mimicking influencers, hoping the algorithm will give them similar treatment. But what such AI researchers don't realize is, this is a race to the bottom. Random influencers will keep lowering the bottom.
Leshem (Legend) Choshen π€π€@LChoshen@deliprao I had an interesting read following your comment. Yes seems like similarity is still a thing. Apparently, you can copy behavior, which will copy who sees it (how similar of an audience are researchers and AI influencer audiences? quite similar?)
looking deeper, current social media is challenging to be who you are. Will you be real, show your authentic self, rise above metrics, and be okay with who recognizes your work and who doesn't? Or will you gradually dissolve into the mode? That's what the algorithm is testing. BTW, this is not just social media. Algorithms are everywhere. https://deliprao.substack.com/p/the-great-ai-weirding
Leshem (Legend) Choshen π€π€@LChoshen@deliprao I'm selective at what I adopt (MIT is never my first lineπ‘... despite being at MIT ATM) FYI lurkers: there are good things to adopt as well. I quite like the down-to-earth and the bottom line of social media (here is what matters to most, details in another format git\paper)
@deliprao I had an interesting read following your comment. Yes seems like similarity is still a thing. Apparently, you can copy behavior, which will copy who sees it (how similar of an audience are researchers and AI influencer audiences? quite similar?)
Delip Rao e/Ο@deliprao@LChoshen embeddings based similarity in recommendations is a real thing.
@deliprao It doesn't change the fact that you can choose what to be similar to; there are many options. Also, I wonder, do you use this knowledge somehow? Or think of ways to use this knowledge (without decreasing value as in the tweet mentioned)
Leshem (Legend) Choshen π€π€@LChoshen@deliprao I had an interesting read following your comment. Yes seems like similarity is still a thing. Apparently, you can copy behavior, which will copy who sees it (how similar of an audience are researchers and AI influencer audiences? quite similar?)
@deliprao It is also a race they can't win. What they have over influencers is exactly their depth of understanding, reliability etc. Influencers are making every effort to be good at surface-level cues.
Delip Rao e/Ο@delipraoyes, the problem is AI researchers are coveting a large overlapping subset of the audience AI influencers are. In the past, that was not a thing. AI researchers cared about other AI researcher and traditional media attention. With death of trad media, incentives beyond traditional academic incentives, current AI researchers are mimicking influencers, hoping the algorithm will give them similar treatment. But what such AI researchers don't realize is, this is a race to the bottom. Random influencers will keep lowering the bottom.