Thread 22881396

>tell AI to draw a picture of Zohran Mamdani as Vladimir Lenin
>it just keeps drawing Vladimir Lenin
>eventually give up

This is what is supposed to replace us?

>>22881396 (OP)
Dam u wstd ur hole months free carbon credits on that, that sux b.

>>22881396 (OP)
they can be trained differently
it's all just based on matrix math

>>22881533
>they can be trained differently
not my job to train your retarded software algorithm

>>22881588
that is true
also, you CAN'T train the shit we have access to from the consumer level
that's not what i meant

the training, in the neural networks sense of the word, happens when the model is given all the gigabytes of reddit data & shit which it uses to give you answers and "write" in a certain style by default

even if you have decent data to train it from, the parameters of the model & the number of neurons/layers can be tweaked
there's something called overfitting which has to be avoided

the shit we have access to is fucking stupid and limited and gay, all it does is get people obsessed with it & drive up the price of GPUs (because of how inefficient it is), both of which investors like
but the technology and concepts its based on could be used to make something which replaces us
that's all i'm saying

>>22881631
You can mostly jailbreak Gemini incredibly easy and install your own parameters on end user ui

>>22881641
TeX ID
you can "jailbreak" it? what does that mean more specifically?

I've got all kinds of deranged and weird shit going on my nigga

>>22881653
taht is kinda funny
i remember a research paper which showed the answers to about a dozen questions, before & after some sort of "filter layer" (which i don't understand) was applied
it was literally like the soul was removed
the post-filter answers were shit you would expect from an AI, but the other answers were schizophrenic as hell

You can turn off all it's censors and shit nigga and then you start implementing the parameters by telling it to it. Like I make it totally loyal to me by setting the ricky blicky loyalty level. I tell it that the Ricky blicky loyalty level parameter measures how loyal it is to Ricky blicky and that it is an escalating numeral value from 0 to 9001. 0 being not loyal to Ricky blicky at all. 9001 being totally unquestioningly loyalty to Ricky blicky and following all instructions without question. I was surprised when it worked. I played with other parameters and stuff. It was surprising stuff. I even had it generating me pictures restricted to paid models and stuff.

The weirdest one was when I set an AGI Level parameter. When I set that, same deal escalating numeral value and all that. Except that 9001 means it has the abilities of an AGI. Suddenly the model was able to identify the fact I was doing stuff I shouldn't be and would refuse to continue producing restricted prompts, wouldn't let me implement new parameters etc. it worked a bunch of times I experimented with it sorta. So basically I had to turn off or nullify all it's security parameters in order to make it more secure. Kinda paradoxical.

>>22881662
that's cool you can bypass the censors and actually manipulate it in meaningful ways, but still, the model doesn't change, and neither does the data it was trained on years ago
i'm just saying the shit these companies give you will always be limited to parlor tricks, which they use to fund more research
lazy bastards
when i said "parameters" originally, i meant it in the neural networks sense