Conversation

Notices

1. Woggy's Zeonic Frolicks (washedoutgundampilot@poa.st)'s status on Saturday, 29-Jul-2023 03:44:10 JST Woggy's Zeonic Frolicks

   How in the hell did AI figure out the Neko Arc voice so well? It’s baffling. There shouldn’t be enough samples for it, and the little flavor bits should be too randomized

   • AWIVR (awivr@poa.st)'s status on Saturday, 29-Jul-2023 03:46:45 JST AWIVR

     in reply to
     @WashedOutGundamPilot you can feed AI generated voice lines back into the AI an infinite number of times.
   • Woggy's Zeonic Frolicks (washedoutgundampilot@poa.st)'s status on Saturday, 29-Jul-2023 03:46:45 JST Woggy's Zeonic Frolicks

     in reply to

     • AWIVR

     @AWIVR How does it learn new patterns though? You’d have to trim all the noise as it keeps branching out without enough original permutations

   • salty_apollyon (rlier23@bae.st)'s status on Saturday, 29-Jul-2023 03:49:23 JST salty_apollyon

     in reply to
     @WashedOutGundamPilot it's not that the there are many samples, it's that there are enough autist who want this outcome so they all put their effort until they succeded
     
     at least that's what i think
     Woggy's Zeonic Frolicks likes this.
   • AWIVR (awivr@poa.st)'s status on Saturday, 29-Jul-2023 03:59:09 JST AWIVR

     in reply to

     • nukie
     @nukie @WashedOutGundamPilot it's counter-intuitive but this is actually how AI is trained in the first place. With a small amount of data as an input you can develop a discriminator that is good enough to tell good inputs from bad ones, then it feeds back into itself hundreds of thousands of times (epochs) until the model is refined enough to give a consistant desirable output.
   • Woggy's Zeonic Frolicks (washedoutgundampilot@poa.st)'s status on Saturday, 29-Jul-2023 03:59:09 JST Woggy's Zeonic Frolicks

     in reply to

     • AWIVR
     • nukie

     @AWIVR @nukie Makes enough sense, it just doesn’t compute how they’d be able to do it without significant man hours in trimming the cancerous cells as you exponentially cultivate fresh flesh

   • nukie (nukie@boymoder.biz)'s status on Saturday, 29-Jul-2023 03:59:11 JST nukie

     in reply to

     • AWIVR
     @AWIVR @WashedOutGundamPilot yeah but that usually results in horrible distortions, no?
   • AWIVR (awivr@poa.st)'s status on Saturday, 29-Jul-2023 04:56:09 JST AWIVR

     in reply to

     • nukie
     @nukie @WashedOutGundamPilot I'm gonna prooooompt
     Woggy's Zeonic Frolicks likes this.
   • AWIVR (awivr@poa.st)'s status on Saturday, 29-Jul-2023 04:56:10 JST AWIVR

     in reply to

     • nukie
     @WashedOutGundamPilot @nukie it works the same with image generation, with a small amount of training data I can tweak an already existing model to output only my character.
   • nukie (nukie@boymoder.biz)'s status on Saturday, 29-Jul-2023 04:56:10 JST nukie

     in reply to

     • AWIVR
     @AWIVR @WashedOutGundamPilot ah, I see you're a proompter. Very nice
   • VidMasterEon (vidmastereon@poa.st)'s status on Saturday, 29-Jul-2023 05:19:55 JST VidMasterEon

     in reply to

     • AWIVR
     @WashedOutGundamPilot @AWIVR If i understand it correctly you just keep iteratively using an encourage discourage model a lot of times
     >make it say a ton of lines
     >discard the stuff that doesnt sound right and feed the stuff that does back into the algorithm
     >repeat until you get the desired result
     It is like training a dog
   • Woggy's Zeonic Frolicks (washedoutgundampilot@poa.st)'s status on Saturday, 29-Jul-2023 05:19:55 JST Woggy's Zeonic Frolicks

     in reply to

     • VidMasterEon
     • AWIVR

     @VidMasterEon @AWIVR But that takes time, because a human has to keep doing that manually and judging it. Given the terrible taste people have, I’m amazed it got so good so quick, instead of getting mediocre quickly