@cmconseils can confirm i downloaded a weird amount of music from a single digit number of people on discord who might no longer wish to be associated with that music
sylvie@chitter.xyz
@sylvie@chitter.xyz
Posts
-
This post did not contain any content. -
Google Chrome doing more absolutely horrid default surveillance fyihttps://bsky.app/profile/meijer.ws/post/3mgevpgqgz22h -
again and again... there is NO SUCH THING, no such thing, NO SUCH THING as automation as presented by AI companies!@olivia thanks for the wiki walk
> Common forms of reward hacking in LLMs include [...] sycophancy, where the model agrees with false user statements rather than giving true information; and sophistication bias, where the model provides false information in a convincing manner. Wen et al. (2024) shows that reinforcement learning from human feedback can make the outputs of large language models more persuasive to human evaluators, even if they are factually incorrect, which they termed "U-Sophistry" (unintended sophistry).
lovely /s