Post

I have a pocket rant about this that I call the "vanilla effect". It starts with some observations I had about rotten tomatoes.

Basically, Rotten Tom's Tomatometer stopped "working" for me at one point. I watched like 3+ 100% movies in a row over a year and they were all boring as hell, for me. Sure enough though, the flicks were highly regarded. So, what was going on?

I eventually realized it was this: the films weren't great. They were safe. No critic watched them and thought, "wow, I can't recommend this to people". They were vanilla.

I'm not criticizing the films or vanilla (it's my favorite flavor), it's not derogatory. Vanilla's an incredibly exotic thing. Extracting it is a culinary milestone. BUT - it's very safe. You can put it in most sweet things and it works. So, it's easy to recommend.

The problem here is that "easy to recommend" means that it's inoffensive, not that it's ideal. And that's why recommendation algorithms suck - they're primed to fulfill success criteria, and that means they hedge against failure.

Or in other words, recommenders pick safe recs. That's also why Spotify sucks so hard these days - at one point there was a team of data nerds who tagged and linked songs with stuff only humans would. They're all gone, replaced by automation.

So that means that the old algorithms that would serve you things like Discover Weekly, and the year in review feature, the actual strange human tastes that guided those into the realm of actual utility have been replaced by an autocomplete robot.

The old Spotify algorithm kicked ass but it was tough to optimize and involved a lot of human parts. Here's my assertion: any recommendation system which a human will get value out of, long-term, must be curated by humans.