Why Would Meta Download So Much Porn?
theatlantic.comI'm glad Zuckerberg appreciates my homemade porn. Nobody else seems to.
Because of the size?
You ever try downloading a video file with only one seed?
The author seems surprised that the downloaded content is only "loosely connected," like a blog post about selling a house and a porn flick about a realtor. But that just proves that no humans were involved in scoping the content at all.
AI training data was vacuumed up without any consideration for the source or even the content itself. The fundamental way LLM models are built invites you to do just that. To get to the biggest "L" (for "large") you must suck up as much content as possible.
They brought it on themselves, by design, and are only now finding out what they actually ingested.
Where are you getting the "blog post about selling a house" from? The examples given are "Bajillion Dollar Properties S01E07" (which is titled "Happy Ending" and involves real estate brokers getting massages: https://en.wikipedia.org/wiki/Bajillion_Dollar_Propertie%24#... ) and "First Time Home Buyer Anal Fantasy" which I assume is actual porn. That seems much more like a human expanding their porn collection at work (and perhaps being disappointed that the first one they downloaded is just a comedy) than an algorithmic crawler.
Meta’s not giving up on VR!
to train models to detect and remove porn from their platforms
thread closed
My experience with image gen makes me think that all of the models include porn in their datasets.