macrocheira

kitty cat extraordinare

soweli Sawa -- ıııᕈ 「^.^ ˍ∧ˍ \O/ a̍̍̍」 
also known as "radiotomatosauce99" on youtube
I make silly little songs sometimes and I aspire to make silly little games :)
18 year old white autistic bi therian transgender woman and/or cat
my 18+ alt account is @soweli-sawa


macrocheira
@macrocheira

machine learning is so cool and it sucks that it's become part of the corporate slop cycle because there's just so many fascinating things that machine learning can do and even so many fascinating things about what machine learning is doing right now


macrocheira
@macrocheira

the idea that chatgpt, which functionally is just turbo predictive text, can do things like coding even half decently is astonishing. language learning models being used generally like this is running on the philosophy of "the best way to sound correct is by actually being correct; therefore if you get good enough at sounding correct, you'll end up inevitably being correct" which sounds like nonsense, but it actually almost works, which is incredible I think!


macrocheira
@macrocheira

machine learning is also I think a fascinating way of analyzing a dataset, the machine has no human biases aside from those present within it's dataset, and therefore any biases it gains must be somehow reflective of the dataset. machine learning models will exacerbate issues within their data that someone who's not a machine might have been able to account for, which is why they can't really be trusted to do anything super important, but in the same vein if used for analysis it can be very insightful since it will make biases much more prominent. it can also be really funny, like that one thing where someone was using an image generation ai to make images of anime girls watching sports, but inexplicably every time they were a jets fan they were always crying. it can't have just made up the connection between the jets and crying out of nothing, so the dataset must have some bias showing that jets fans are more often unhappy, which to me is a really roundabout and funny insult


macrocheira
@macrocheira

and also I love seeing machine learning models for things like making a simulated person walk because sometimes they end up doing silly things you wouldn't expect in a way that kind of pokes fun at the limits of the simulation itself. you didn't want it to fold in on itself and then launch itself magically? well you probably should have told it not to, silly


You must log in to comment.

in reply to @macrocheira's post:

One of my absolute favorite stories about this sort of thing is when some researchers trained a neural network to play NES games, heavily penalizing it for getting a game over state. It did pretty good on most games, but just couldn't figure out Tetris. Since there was no win state, it would eventually lose - so it learned to pause the game immediately and never do anything else.

Earlier text generators like the one used in Talk To Transformer / Inferkit (rip) were prone to making very off-the-wall texts, which were often either ominous or kinda funny. I hate the chatgpt format of interaction, just give me the pure GPT that'll try continuing a text without pretending to be a real person and without all its bizarre filters. "As a neural network, I can't..." Yeah? Suck my balls.