What do I think about this?

Well, going off of what I’m going to call ‘internal heuristics reinforced by 35 years of raw data’:

The author is very likely female and Chinese.

They use more harsh, accusatory, and direct language when describing men, and they use more soft, explanatory, and indirect language when describing women.

The various grammatical oddities sound extremely similar to ESL, native Mandarin and Cantonese speakers I’ve known personally.

Ok, from my ‘career as a data analyst’ perspective?

AB Testing

…

This is the only actually useful part from a ‘how do you actually make a good dating app’ perspective. It is difficult to AB test the algorithms of the feed because the change usually has an impact on the user that gets a specific variant but also users that he likes that are on any variant. Other than that, you can gain 10 points of retention just with AB testing and finding the best things that work.

It is actually pretty easy to create a dating app that works in terms of retention if you use extensive AB testing.

Yes, yep, AB testing works.

Too bad the vast majority of the analysis is devotes to explaining how things don’t work or aren’t that impactful or at best, prevent retention from falling.

The word ‘retention’ appears 72 times in this post, and that blurb I quoted there is literally the only time it is mentioned within a context of doing something to change the algorithm, the underlying nature of the app, that improves retention, by a specific amount, across the board.

That means AB testing is the actual secret sauce, and well, there aren’t any details because then the sauce wouldn’t be secret any more.

I have some background in economics.

In econ, specifically hedonics, attempting to determine a consumers actual preference for one basket of goods vs another, you’ve got the core concept of cardinal preferences and ordinal preferences.

I’ve also got a background in poli sci.

The same basic concept underlies ranked choice vs first past the post voting.

Long and short of it is: To get a result that does a far better job of actually understanding preferences, and matching them with outcomes… you have to de-abstractify decisions and make people actually think about them.

An example in a dating app would be: If you match on someone who has something you’ve indicated as a red flag, but pass on someone who doesn’t… you tell the user what they did, and ask the user why they did that.

The ‘good’ way to proceed with this would be to inform the user when they are being inconsistent, to actually help them figure out what they actually want, explain their inconsistencies to them, have a way of suggesting

The ‘bad’ way to proceed with this would be just track this kind of ‘hypocrisy’ score in the background, and not help the user become less of a hypocrite. Then, you sprinkle (or firehose!) peoples feeds with hypocrites who have a decent enough chance of matching, but you know have a high likelihood of having a it not working out after some weeks or months, thus creating a perpetual soul grinding machine that keeps you coming back.