stopthatgirl7 to Technology@lemmy.world • 11 months agoEverything Made By an AI Is In the Public Domaindoctorow.medium.comexternal-linkmessage-square138fedilinkarrow-up1664arrow-down132file-text
arrow-up1632arrow-down1external-linkEverything Made By an AI Is In the Public Domaindoctorow.medium.comstopthatgirl7 to Technology@lemmy.world • 11 months agomessage-square138fedilinkfile-text
minus-square@illusoryMechanist@lemmy.worldlinkfedilinkEnglish3•11 months agoDoes this mean we can, indeed, train AI on ChatGPT/other AI outputs?
minus-square@Natanael@slrpnk.netlinkfedilinkEnglish9•11 months agoFYI there’s studies showing that repeatedly training ML on ML outputs degrades quality sharply after 4-5 cycles
minus-square@Prager_U@lemmy.worldlinkfedilinkEnglish0•11 months agoThat is a used method for training other LLMs, I believe OpenAssistant did this.
minus-square@assassin_aragorn@lemmy.worldlinkfedilinkEnglish0•11 months agoDoesn’t that break the models?
minus-square@Natanael@slrpnk.netlinkfedilinkEnglish2•11 months agoIt does after 4-5 cycles yeah, built in algorithmic biases gets amplified to a degree that it breaks the model
Does this mean we can, indeed, train AI on ChatGPT/other AI outputs?
FYI there’s studies showing that repeatedly training ML on ML outputs degrades quality sharply after 4-5 cycles
That is a used method for training other LLMs, I believe OpenAssistant did this.
Doesn’t that break the models?
It does after 4-5 cycles yeah, built in algorithmic biases gets amplified to a degree that it breaks the model