Living Dan-gerously

Started a new blog.

Once, I encountered the funny story of an AI image descriptor with a sheep obsession. It had been trained on pictures of fields of sheep. Therefore, it tagged anything in a field as 'sheep', including an empty field, because they work on statistical probability. Therefore, it thinks "ah, a field! there's probably a sheep here." (It's a bit more complicated but basically that.) It also couldn't recognise sheep in places that weren't fields, such as petrol stations or barns. [cont]

But wait, it’s even weirder than that!

…. oh my fucking god they actually are using open source software. They’re using a fucking one-layer unidirectional bicategory tag-trained neural network. This will never work. Literally, it will never work. There’s just not enough algorithmic complexity to do what they’re asking of it. I bet you I could prove on a mathematical level that this joke of a neural net fundamentally lacks the abstraction necessary to do its job.

Also, just to make this clear, Tumblr lied. I mean, we already know this, but I mean they liiiieeeeed. All that stuff they promised about what would or would not be censored? That cannot be delivered on with a system this simple. Nude classical sculptures, political protests, male-presenting nipples (really Tumblr?), nude art outside the context of sex, all that? You cannot train a bicategory one-layer neural network to exclude those things. It cannot be done. Tumblr never intended for those things to actually be permitted, they were just lying. Because the system they have cannot actually do what they said it would and never will be able to.