I am seeing companies like Dippy AI (character ai competitor for gen-z) scale from 0->10M+ users rapidly without many people noticing.
So, 20% of users who send a message just retain forever.
50% leave within 1 week
75% leave within 5-10 weeks
85% leave within a year
Also, I'm not particularly inclined to trust a chart which can't even maintain standard intervals for its tick marks, let alone failing to start its Y axis at 0. Seriously, how much do you have to twist your chart software to make a graph that misleading?
Is user retention an ergodic process? Clearly no. From economic first principles, we know elasticity of demand is going to depend upon substitutability and disposable income of the user base. Both of these can and do change over time. From empirical results elsewhere, we can simply look at products that have existed for more than a single year and observe that user retention rates in 1995 did not always match retention rates for the same product in 2002.
Here we get measurement of a single cohort that has actually reached 50 weeks. Are they representative of a typical cohort? We have no way to know. And again, fair, this is the best the company can possibly do. They can't create data that doesn't exist and they can't draw conclusions that have no hard evidence for or against. But you don't need to draw unwarranted conclusions, either. It's also 15% at 50 weeks retention in that first cohort, not 20% at 52 weeks. Does that mean those 15% of users will stick around forever? We have no idea. Sort of. We actually know for certain they won't because they're mortal and will die. Setting that aside, might they at least become lifelong users? I suppose it's possible, but even real-world friendships rarely last a lifetime. Will the next cohort at least still show 15% at 50 weeks? We yet again have no idea.
Did you get your MBA from the back of a mail order catalog?
There may be high retention use cases that find ways to serve user needs without embarrassing them.
Civit.ai, the primary competitor to huggingface (i.e. github for AI) for diffusion models, might as well be a brothel with how horny it is. Over half of all "low rank adapters/Loras" every created are most likely NSFW in nature.