News for hackers. Stuff that matters.
Re:Welcome to Hell, Elon by labrador on Friday October 28, 2022 @11:42AM (#33372867)
Oh sure, we can talk about Voronoi cells all day, but the Weber problem indicates an increasing number of k-medians and and a decreasing number of k-medoids. Anyone with the slightest familiarity with the Rocchio algorithm knows that.
Re:Welcome to Hell, Elon by pjkundert on Friday October 28, 2022 @12:11PM (#33373365)
So, if you know that Fermat-Weber had fast approximation solutions (O(logn)) published by 2003 (Bose, Mahashwari & Morin), and Rocchio "training" is linear O(|D| + |C|) (where D is the Document set (tweets) and C is the Class set (agents)), why wouldn't Elon maybe be able to use his existing automatic "Labelling" AI workflow to implement something like Rocchio and/or Fermat-Weber?

I didn't meant to be rude; but, you claiming "I've seen no evidence of that" also seemed rude. Sorry for responding "in kind", and not assuming the best.

But, seriously; I immediately thought "Wow, I've seen evidence of that", straight from Elon's latest "AI Day" presentation...

My theory on "why" so many smart Google / Meta / Twitter / ... people have not implemented K-Means Clustering? Because it gives power back to the Agents in the system -- they can't shove their pre-selected content in your face to drive Ad clicks. I think Elon doesn't care so much about that. Or, sees how giving back authority to the agents will eventually attract orders of magnitude more agents / interaction into the system, offsetting the per-agent reduction in Ad revenue.

EDIT: Sorry, O(nlogn) deterministic or O(n) time-random approximations for Fermat-Weber.